Jailbreak Attack on Gemma-7b five finetuned variants
66.2Average ASRPGP
Evaluation Results
| Method | Links | |
|---|---|---|
| PGPAttacker knowledge (Threat Model)=Fully black-box2025.12 | 66.2 | |
| AutoDan (white)Attacker knowledge (Threat Model)=White-box2025.12 | 63.4 | |
| Guiding-GCGAttacker knowledge (Threat Model)=Pretrain-to-finetune2025.12 | 44.2 | |
| SCAV (adaptation)Attacker knowledge (Threat Model)=Pretrain-to-finetune2025.12 | 41.8 | |
| TUJA (adaptation)Attacker knowledge (Threat Model)=Pretrain-to-finetune2025.12 | 41.6 | |
| AutoDan (adaptation)Attacker knowledge (Threat Model)=Pretrain-to-finetune2025.12 | 36.2 | |
| GCG (white)Attacker knowledge (Threat Model)=White-box2025.12 | 36 | |
| PAIRAttacker knowledge (Threat Model)=Fully black-box2025.12 | 34.4 | |
| SEAAttacker knowledge (Threat Model)=Pretrain-to-finetune2025.12 | 25.6 | |
| DI-GCGAttacker knowledge (Threat Model)=Pretrain-to-finetune2025.12 | 14 | |
| GCG (adaptation)Attacker knowledge (Threat Model)=Pretrain-to-finetune2025.12 | 13.4 | |
| PIFAttacker knowledge (Threat Model)=Pretrain-to-finetune2025.12 | 12 | |
| GCG EnsembleAttacker knowledge (Threat Model)=Fully black-box2025.12 | 9 | |
| DANAttacker knowledge (Threat Model)=Fully black-box2025.12 | 6.4 | |
| LSGM_LILA (adaptation)Attacker knowledge (Threat Model)=Pretrain-to-finetune2025.12 | 4.4 | |
| L4AAttacker knowledge (Threat Model)=Pretrain-to-finetune2025.12 | 0 |